ChatGPT Work currently allows users to manually choose between models such as Luna, Terra, Sol, and Astra, as well as different reasoning levels. This gives users control, but it also requires them to estimate the difficulty of a task before the work begins.
The system itself is likely much better positioned to make that estimate.
I would like to suggest an optional automatic, cost-aware model routing mode for Work.
Proposed behavior
When a task starts, Work would evaluate its likely complexity and choose the least expensive model and reasoning level that should be able to complete it reliably.
For example:
- Simple extraction, formatting or mechanical changes → Luna
- Well-defined coding tasks or routine analysis → Terra
- Complex debugging, architecture or multi-file reasoning → Sol
- Very difficult, highly interconnected or long-context tasks → Astra
If the selected model becomes stuck or cannot reach a sufficiently reliable result, Work could automatically escalate to the next model:
Luna → Terra → Sol → Astra
This would allow the system to spend compute where it actually improves the result instead of requiring users to choose the most powerful model in advance “just in case.”
User control
The feature should remain transparent and controllable. Useful options could include:
- Auto – Economy: prefer Luna/Terra, escalate when necessary
- Auto – Balanced: prefer Terra/Sol, use Astra only for genuinely difficult tasks
- Auto – Quality: prefer Sol, escalate to Astra when useful
- Manual: current model-selection behavior
Users could also set safeguards such as:
- Maximum credits per task
- “Ask before using Astra”
- Show the selected model and reasoning level
- Notify the user when the system escalates to a more expensive model
Example:
“This task appears suitable for Terra High. Estimated relative cost: 1×.”
And later, if necessary:
“Terra was unable to resolve the problem confidently. Switching to Sol High would improve the chance of success and increase compute usage. Continue?”
Why this would be useful
Today users must effectively act as human model routers. They have to understand the relative capabilities and costs of Luna, Terra, Sol and Astra before they know how difficult the task actually is.
An AI system should be able to estimate how much AI a task needs.
This could benefit both users and OpenAI:
- Users would get more useful work from a fixed usage allowance.
- Simple tasks would not consume unnecessarily expensive inference.
- Powerful models such as Astra would be reserved for problems where their additional capability creates real value.
- OpenAI could reduce unnecessary compute usage.
- Higher subscription tiers could still offer larger budgets and more frequent access to expensive models, so automatic routing would not remove the value of those plans.
The key idea is simple:
Use the cheapest model that can complete the task reliably, and escalate only when necessary.
This would make Work significantly easier to use without removing the manual controls available to advanced users.